beyond machine learning...circulation what information goes into creating well offset ... bit record...
TRANSCRIPT
AI & COGNITIVE SYSTEMS FORUM | 29-31 OCTOBER 2019
BEYOND MACHINE LEARNING
DEVELOPING TRULY INTELLIGENT SYSTEMS BY EMBEDDING HUMAN EXPERTISE TO DIGITAL TWIN
BEYOND MACHINE LEARNINGTERESA TUNG
Copyright © 2017 Accenture. All rights reserved. 3
BEYOND MACHINE LEARNING
DIGITAL TWIN
DATA
DIGITIAL TWIN
OLDWORK PROCESS
NEW / MODIFIEDWORK PROCESS
CREATE VALUE
• New markets• Time to market• Regulatory compliance• HSE• Production quality• Yield• OPEX
2D Drawings
Estimates
Photos
VideosProcess knowldege
3D Point Clouds
3D Meshes
Commercial data
Certificates
IoT Sensor data
3D Models
Transactions
Analytics & AI
PLM
Supply Chain
IoT
MESMOM
CAD
DIGITAL TWIN IS A DIGITAL MODEL OF A PHYSICAL OBJECT, SYSTEM OR PROCESS THAT EXISTS (OR COULD EXIST)
Copyright © 2017 Accenture. All rights reserved. 4
BEYOND MACHINE LEARNING
EXAMPLE: DIGITAL TWIN FOR OIL WELL
DIGITIAL TWIN
Well DataField dataEWRDDRAFECosts
ProductionDecline CurvesHistorianReal-Time OperationsComplianceERP
ML supports Continuous Improvement Across
Multidisciplinary teams
ML generates Optimal Well Design and Execution plan
AI-driven Drilling & Completion
Predictive Analytics on Integrated Digital Supply Chain
Real-Time Analytics for Wells Optimization Center
WELL DESIGN DRILLING & COMPLETION
Analytics & AI
PLM
Supply Chain
IoT
MESMOM
CAD
Copyright © 2017 Accenture. All rights reserved. 5
BEYOND MACHINE LEARNING
+HUMAN EXPERTISE
LIFECYCLE OF THING / SYSTEM
DTLeverage past performance results to generate designrequirements DT
Reconcile as-designed and as-built through manufacturing build and commisioning
DT
Leverage past designs to generate engineering detailed design to fulfill requirements
DTApply cross system lessons with instance-specific needs to optimize operations
ADDS THE CONTEXT FOR DIFFERENT VIEWS OF THE DIGITAL TWIN
What results are interesting? E.g., high productivity, loss circulation
What information goes into creating well offset study?
What measurements are used? What is acceptable variance?
What is the insight from the specific pattern of events? What is the next-best action?
Copyright © 2017 Accenture. All rights reserved. 6
BEYOND MACHINE LEARNING
EXAMPLE: +HUMAN EXPERTISE IN SEARCHEXAMPLE: INTERNET SEARCH FOR ACCENTURE
Copyright © 2017 Accenture. All rights reserved. 7
BEYOND MACHINE LEARNING
KNOWLEDGE GRAPH
Extract Information Contextualize Information Assemble Information
Painting
Is a
Mona Lisala Joconde a’Washington
Person
talks about was created by
Is a
Leonardo Da’Vinci
was created by
EXAMPLE: INTERNET SEARCH
Copyright © 2017 Accenture. All rights reserved. 8
BEYOND MACHINE LEARNING
DIGITAL TWIN + HUMAN EXPERIENCE
CREATE A DESIGN IDENTIFY A ROOT CAUSE REDEFINE A PROCESS
Fall, Slip Notice, Walk
Fall, Trip,
Puncture
Floor,
Slip, Leak
Water,
Oil,
Leak
Floor,
Water,
Slip
Pole, Wire,
Crew
Work,
Area,
Hazard
Electric,
Pole,
Crew
Wasp, Snakes
“Safety event”
Copyright © 2017 Accenture. All rights reserved. 9
BEYOND MACHINE LEARNING
MODEL HUMAN REASONING
Start with the questions
• How should we plan our next well?
• What are similar offset wells?
• What are events I should be aware of?
• What sort of impact do they have?
• How often does it happen?
• Under what conditions?
Example Model Schema
Copyright © 2017 Accenture. All rights reserved. 10
BEYOND MACHINE LEARNING
POPULATE THE GRAPH
Non-Exhaustive
Data Sources
DDRBit recordNOOP &
AFE
FWR FDPHIS
Report
Post Drilling
Report
Final
Geologic
al Report
Final
Cement
Report
Mud LogsRushmore
Final
DD/MWD
Report
Structured Unstructured Legend:
Drilling Data TypeDrilling Report
Drilling Databases
• Well Architecture (stick chart)
• Actual Mud weight ranges, cement
type & quality
• Downhole related problems
• NPT data
• Lessons Learnt/After Action
Review/TL Performance
• Well Depth
• Well Cost & Days
• Rig type
• Time and date of historical well
• Geological & Lithological Prognosis
• Any relevant regional info/issues
• Sidetracking method & info
…
Example from Drilling Report:
Skid out cantilever 50 ft to
transom aftwards, 2 ft transverse
portwards to well center and
secure same.
300m
Example Well Design Schematic
200m
Natural Language
Processing• Regular Expressions
• Dictionary-based approaches
• Pattern-based approaches
• Machine Learning
• Other combinations
Image Processing• Image Recognition
• Image Detection
• Image Classification
• Edge Detection
• Corner Detection
Relational Data Processing• Schema Parsing
• Entity Relationship Diagram to Hierarchy
Parsing
• Data Normalization/ Denormalization
Extract from Original System
Copyright © 2017 Accenture. All rights reserved. 11
BEYOND MACHINE LEARNING
WELL ADVISOR: OFFSET STUDY
1
Identify population of wells
that meet the search criteria
Copyright © 2017 Accenture. All rights reserved. 12
BEYOND MACHINE LEARNING
WELL ADVISOR: OFFSET STUDY
Copyright © 2019 Accenture. All rights reserved.
2
Generate a
unique profile
based on the
population
3
Recommend
similar wells
to the profile or
identify outliers
Measurements
Well Parts Materials
Rig PartsOperations
Copyright © 2017 Accenture. All rights reserved. 13
BEYOND MACHINE LEARNING
WELL ADVISOR: OFFSET STUDY
4
Identify
commonalities
or abnormalities
in the curated
population
5
Track remediation and explain
consequences or commonalities.
Event
Type
Operation
Type 1
Operation
Type 2
Event
Type
Corrective
Action 1
Operation
Type 2
Corrective
Action 2Remediation
Copyright © 2017 Accenture. All rights reserved. 14
BEYOND MACHINE LEARNING
YOU NEED HUMAN EXPERIENCE TO…
+HUMAN EXPERIENCE
• Configure and Apply Digital Twin
• Handle Domain-specific models, processes, regulations
• Augment Gaps in Data
Populate with Data
Exploit Knowledge
Capture & Curate Human Reasoning
AI & COGNITIVE SYSTEMS FORUM | 29-31 OCTOBER 2019
DIGITALIZING INDUSTRIES
COME JOIN US!
SAVE THE DATE
29 - 31 October 2019FOR MORE INFO VISIT
www.iotsworldcongress.comFOLLOW US
#IOTSWC19